Samples were obtained from 76 patients with PDAC who underwent surgical resection from 2013 to 2024 at the Queen Elizabeth II Health Sciences Centre (Nova Scotia, CAN). This study was approved by the Nova Scotia Health institutional research ethics board (#1023467 & #1026628). Patients who received neoadjuvant therapy were excluded. Patient information is available in Supplementary Table S1.

Patient samples were formalin-fixed, paraffin-embedded (FFPE) and assembled into tissue microarrays (TMA). To mitigate intra-tumoural variability, each patient was represented by duplicate 3 mm cores selected from regions containing >30% tumour cells, as identified by board-certified pathologist (TA).

Opal mIF (Akoya Biotechnology, Marlborough, USA) was performed as previously described (16). Briefly, 4 µm sections were prepared from the TMAs, deparaffinized in Xylene (Sigma-Aldrich) (3 x 10 min) and rehydrated by incubating with decreasing concentrations of ethanol for 10 min each (100%, 100%, 95%, 70%). The slides were then fixed in 10% neutral buffered formalin (VWR) for 20 min.

The slides underwent serial cycles of antigen retrieval, primary staining, and tyramide signal amplification to add each antibody and fluorophore one by one. For antigen retrieval, slides were microwaved in Tris-EDTA buffer (pH9) at 1000W for 2 min, then 200W for 15 min. The slides were cooled on ice for 10 min and rinsed in deionized water and Tris-buffered saline with Tween-20 (TBS-T). Slides were thereafter blocked in 1x antibody diluent/block (Akoya Biosciences) for 10 min, the excess block was removed, and the slides were stained with diluted primary antibody for 45 min. After TBS-T rinsing, slides were incubated in secondary anti-mouse plus rabbit antibodies conjugated to horseradish peroxidase (HRP) for 10 min. Signal amplification was completed via HRP-driven tyramide activation and fluorophore deposition and tissue binding (Opal 6-Plex Manual Detection Kit, and Opal 480 Reagent Pack, Akoya Biosciences). Opal 780 staining (Opal 780 Reagent Pack, Akoya Biosciences) required use of digoxygenin labeling prior to tyramide signal amplification (10 min, TSA Plus DIG, Akoya Biosciences). The slides were rinsed in TBS-T and subsequently incubated in the corresponding Opal fluorophore diluted in Opal amplification diluent for 10 min (Opal 480, 520, 540, 570, 620, 650, and 690 fluorophores) or 1h (Opal 780). After the final staining step, slides were counterstained with Spectral DAPI (ThermoFisher) for 10 min, rinsed, and mounted in Mowiol-based mounting medium (Mowiol, glycerol, deionized H₂O, TrisHCl (pH 8.5); Sigma-Aldrich)). Antibodies, Opal fluorophores with which they were used, and staining order are shown in Table 1.

Antibody staining was validated using FFPE tonsil tissue and cell pellets of purified T and NK cells. T and NK cells were isolated from primary peripheral blood mononuclear cells (Canadian Blood Services, Dalhousie REB 2016-3842), centrifuged at 300 x g for 5 min, fixed in 10% neutral buffered formalin, and processed for paraffin embedding. All staining patterns were confirmed by a board-certified pathologist (TA). Representative images from one case and the segmentation pipeline are shown in Figure 1A.

Imaging was completed using a Mantra Quantitative Pathology workstation equipped with after-market Opal 480 and Opal 780 filter cubes. Images were acquired at 20x magnification, with wavelength intervals spanning emission peaks between 420 nm and 780 nm to capture all Opal fluorophores. Up to 15 images per core were captured from tumour-rich regions selected by a board-certified pathologist (TA) to mitigate intratumoral variability. Multispectral analysis performed in InForm Tissue Finder Software (Version 2.4.8, Akoya Biosciences). Spectral reference libraries, generated from monoplex stains for each fluorophore, were used for spectral unmixing to distinguish fluorophore signals and control for autofluorescence. Regions of benign panCK+ staining were omitted and regions of tumour epithelium were confirmed by a board-certified pathologist (TA).

Using machine learning algorithms in the InForm Tissue Finding Software, tissue segmentation was performed on spectrally unmixed images to distinguish epithelial and stromal regions (off-core regions were omitted and could be distinguished from stroma by lack of DAPI staining and tissue autofluorescence). Nuclear segmentation (DAPI) followed, and cell phenotypes were assigned by antibody staining. Phenotype maps with x,y cell coordinates classified each cell with a unique cell ID (Figure 1B). Cell types were classified based on antibody expression profiles: panCK+ as epithelial cells (regardless of other antibody binding), PanCK–CD68+ as macrophages (CD163+/–, CD16a+/–), PanCK–CD68–CD3+ as T cells (CD8+/–, CD16a+/–, CD57+/–), PanCK–CD3–CD68–CD57+ as NK cells (CD16a+/–), PanCK–CD3–CD68–CD57–CD16a+ as CD16a+ cells, and panCK–CD3–CD68–CD20+ as B cells. Cells staining with DAPI only were categorized as “Other”. Antibody binding patterns, cell surface staining, cell segmentation, and phenotyping were validated by a board-certified pathologist (TA).

Quality control involved visual assessment of every image and excluding images with poor tissue quality or non-specific staining. All quality control was performed blinded to clinical outcome data. Of the 76 patients intended for study, images from 73 patients passed staining and quality control and were included in the final analysis.

Our spatial biology workflow was adapted from previously published methods (15, 16) and conducted in Python (Version 3.9.18). X,y coordinates of each cell were used to reconstruct phenotype maps of tumours, and each cell was defined as an index. Cell IDs with their x,y coordinates and assigned phenotypes were grouped by image and patient using Pandas (30) (for data organization and manipulation) and NumPy (31) (for numerical operations). Using the BallTree module from scikit-learn (32), for each index cell, cells (phenotype and count) within a 30 µm radius were identified based on their x,y, coordinates (nearest neighbour search), defining local “groups” (Figure 1C). To identify patterns of immune cell co-localization, which we termed “neighbourhoods”, three images from each patient were randomly selected and combined to create a training set. Using this set, K-means clustering through MiniBatchKMeans algorithm (scikit-learn) was applied to the immune cell populations only, testing between 2 and 10 potential neighbourhoods. The optimal number of neighbourhoods to represent intra-tumour diversity was defined using Davies-Bouldin indexing (scikit-learn) (33).

All remaining images were classified using the trained model to assign each cell to a neighbourhood. To ensure neighborhood definitions reflected immune cell organization (not tumour or other cells), epithelial (PanCK+) and “Other” cells were excluded from the clustering and instead assigned to distinct “PanCK” and “Other” neighborhoods. This approach preserved their spatial position in the radius-based analysis while preventing them from influencing immune neighbourhood structure. Consistency of neighbourhood classification was validated by calculating cosine similarity (scikit-learn), which scored >0.99 between the initial training set and the subsequent classified images. Figures were generated with matplotlib (34) and seaborn (35).

To define tumour regions as intratumoural, peritumoural, and stromal, cell clustering was performed in Python using the DBSCAN (36) and SciPy (37) packages. Epithelial (panCK+) cells were first identified in each sample, and DBSCAN clustering was performed on these cells using a 30 μm radius and a minimum cluster size of five cells. Cells that did not meet this clustering criteria were excluded from defining intratumoural regions. SciPy’s ConvexHull class was used to define the smallest bounding shape (convex hull) for each tumour cluster, with visual confirmation by a pathologist (TA). Based on these boundaries, three regions were defined: (1) Intratumoural: the area formed by each tumour site; (2) Peritumoural: a 30 μm perimeter surrounding each tumour site; or (3) Stromal: the area outside the intratumoural and peritumoural regions. Cell counts for each region were aggregated by radius for each patient, and patients were subsequently classified into high and low groups based on the proportion of each immune phenotype within the defined regions, as described above.

To define which clinical variables were associated with survival and therefore should be used as covariates for downstream multivariate analysis, we performed univariate Cox proportional hazard regressions using five-year OS as the dependent variable in RStudio. Independent variables were tumour stage, tumour grade, age at diagnosis, sex, and lymph node invasion (Supplementary Table S1). Among them, only increased regional lymph node metastasis was significantly associated with shorter five-year OS (p < 0.05).

Univariate Kaplan-Meier and log-rank tests were conducted to assess differences in OS between high and low groups of cell-, neighborhood-, and regional- analyses, and multivariate Cox proportional hazards models were used to identify independent prognostic factors. Co-infiltration between cellular phenotypes were assessed using Spearman’s rank correlation in GraphPad Prism (GraphPad Software, Boston, MA USA, version 10.2.3). Comparisons of immune cell infiltration across clinically defined groups, including tumour stage (T1, T2, T3), tumour grade (1, 2, 3), age (<65, ≥65), sex (male vs female), and regional lymph node metastasis (N0, N1, N2), were conducted using two-tailed Mann-Whiteney U tests in GraphPad Prism, with corrections for false discovery to account for multiple comparisons. Statistical significance was defined as an alpha of 0.05 for all analyses.

Using Opal mIF, we evaluated tumours from a total of 73 patients with PDAC who underwent surgical resection between 2013 and 2024 at the QEII Health Sciences Centre. Forty-five percent of the cohort was female; 55% were male (Supplementary Table S1). Most patients were >65 years old (59%, n=43), and categorized as stage II (33%) or stage III (62%). Moderately differentiated tumours were most common (63%, grade 2), followed by poorly differentiated (23%, grade 3), and well-differentiated tumours (14%, grade 1). Regional lymph node involvement was most commonly present: just 14% had no positive lymph nodes (N0), 60% of patients had 1–3 positive lymph nodes (N1), and 21% having ≥4 nodes (N2). Patients with N0 disease had a lower risk of death compared to N2 (HR: 0.16, 95% CI: 0.04-0.73, p<0.05). Since this was the only clinical variable associated with OS, it was the only one included in downstream multivariate analyses.

The immune microenvironment of PDAC is immunosuppressive and spatially heterogenous, yet remains poorly characterized, with limited knowledge of how it reflects patient prognosis (4, 5). To characterize the immune landscape of PDAC, we quantified immune cell densities (cells/mm²) in both the epithelial and stromal compartments using multiplex IHC and quantitative pathology (Figure 1). Tumours displayed marked immune infiltration with inter- and intra-patient heterogeneity (Figure 2A). Regardless of phenotype, immune cell densities were higher in the stroma than in the epithelial compartment, likely reflecting limited accessibility to the tumour (Figure 2A). Among immune populations, the most abundant were CD16a+ M2-like macrophages (epithelium median: 11.3, min-max: 0-280.1; stroma median: 569.1, min-max: 45.2-4602.0) and CD3+CD8+ T cells (epithelium median: 10.6, min-max: 0-216.1; stroma median: 234.4, min-max: 24.9-1446.0), reflecting substantial infiltration in both compartments (Figure 2A, Supplementary Table S2). To reduce complexity and improve interpretability, populations with a median cohort total density of <1 cell/mm² were grouped into higher-level phenotypes based on shared marker expression (Supplementary Table S2). Specifically, PanCK–CD68–CD3+CD8–CD57+ and PanCK–CD68–CD3+CD8–CD16a+ subsets were amalgamated into the CD3+CD8– T cell group; PanCK–CD68–CD3+CD8+CD57+ and PanCK–CD68–CD3+CD8+CD16a+ subsets were grouped as CD3+CD8+ T cells; and PanCK–CD68–CD3-CD8–CD57+CD16a– and PanCK–CD68–CD3-CD57+CD16a+ subsets were grouped as NK cells.

Previous studies indicated that T cells, B cells, and macrophages, and their specific subsets inform good or bad prognosis in PDAC (38–41), thus we sought to determine whether immune cell densities in PDAC tumour compartments correlated with OS. We used univariate and multivariate Cox regression analyses to compare 5-year OS in patients with immune cell densities categorized as high or low.

By univariate analysis, higher densities of epithelial CD16a+ cells were associated with an increased risk of death (HR: 2.4, 95% CI: 1.4–4.4, p<0.05), while higher densities of CD3+CD8– T cells in both the stroma and total tissue compartments were significantly associated with reduced risk of death (HR: 0.55, 95% CI: 0.31–0.97, p<0.05; HR: 0.42, 95% CI: 0.24–0.75, p<0.05, respectively) (Figure 2B). These associations remained significant after adjusting for regional lymph node involvement by multivariate assessment (epithelial CD16a+: HR: 2.1, 95% CI: 1.1-3.9, p<0.05; stromal CD3+CD8-: HR: 0.47, 95% CI: 0.26-0.85, p<0.05; total CD3+CD8-: HR: 0.43, 95% CI: 0.24-0.80, p<0.05) (Supplementary Table S3). Increased M1-like macrophage density was associated with decreased risk of death in univariate analysis (HR: 0.48, 95% CI: 0.26–0.89, p<0.05), but this association was lost in the multivariate model (HR: 0.55, 95% CI: 0.30-1.0, p=0.07). Thus, immune cell infiltration itself does not suggest improved prognosis: different subsets are variably prognostic and CD3+CD8– T cells may be a marker of improved OS in patients with PDAC.

Immune cells function as a coordinated network. We next hypothesized that the association of CD3+CD8- T cells with improved OS might reflect a role for these cells–which are likely helper T cells–in supporting anti-tumour immune functions, including that of cytotoxic (CD3+CD8+) T cells. Indeed, patients with high infiltration of both CD3+CD8- and CD3+CD8+ T cell subsets exhibit significantly improved OS compared to those low for both (HR: 0.33, 95% CI: 0.12–0.88, p<0.05). While high infiltration of either subset alone trended toward lengthened OS, neither reached statistical significance (CD3+CD8– only: HR: 0.45, 95% CI: 0.17-1.2, p =0.14; CD3+CD8+ only: HR 0.65, 95% CI: 0.14-3.0, p=0.88) (Figure 3A). Notably, high infiltration of CD3+CD8+ T cells alone was uncommon (7% of patients) (Figure 3B). These findings suggest that co-infiltration of helper and cytotoxic T cells might enable local generation of antitumour immunity.

CD8+ cytotoxic T cell killing is contact-dependent, but helper T cells can function from a distance. To evaluate whether compartment-specific localization influences prognosis, we generated a T cell infiltration score (Figure 3C). High T cell infiltration scores were associated with reduced risk of death and improved survival compared to low in both univariate and multivariate Cox regression models (HR: 0.39, 95% CI: 0.16-0.90, p<0.05) and Kaplan-Meier survival analysis (p<0.05) (Figures 3D, E, Supplementary Table S3). Patients with intermediate T cell infiltration scores exhibited a trend towards decreased risk of death and lengthened OS compared to those with low scores, though the difference did not reach statistical significance (HR: 0.46, 95% CI: 0.20-1.0, p=0.066) (Figures 3D, E). These findings suggest that the localization of T cell subsets with respect to tumour cells impact patient prognosis.

PDAC tumours grow as cells around pancreatic ducts, dispersed between dense stroma throughout the pancreas; this is different from other solid tumours where groups of cancerous cells aggregate (21, 22). Dichotomizing tumour versus stroma, as with other tumours, might therefore inadequately capture immune:tumour interactions in PDAC. To refine the assessment of spatial immune localization and its association with survival, we applied DBSCAN-based spatial clustering to define tissue into intratumoural, peritumoural, and distal stromal regions and measure immune cell localization within them (Figure 4A). This approach enabled a more granular evaluation of immune cell positioning at the tumour-stroma interface.

Within the intratumoural region, higher proportions of NK cells were associated with increased risk of death (HR: 2.6, 95% CI: 1.4-4.7, p<0.05) (Figure 4B, Supplementary Table S3). Similarly, increased NK cell infiltration in the peritumoural space was also associated with worse survival (HR: 2.3, 95% CI: 1.2-4.2, p<0.05). In contrast, higher proportions of CD3+CD8-T cells in the peritumoural region were significantly associated with reduced risk of death (HR: 0.55, 95% CI: 0.30-0.99, p<0.05). Within the distal stroma, both CD3+CD8– and CD3+CD8+ T cells were associated with improved OS (HR: 0.53, 95% CI: 0.29-0.94, p<0.05; HR: 0.53, 95% CI: 0.29-0.95, p<0.05). Taken together, these findings suggest that T cell localization, particularly at the tumour-stroma interface, is predictive of prognosis in PDAC.

Although T cell infiltration is often assessed for its prognostic capacity, effective immunity requires cooperation between immune cell subtypes. To determine whether additional cells might interact with T cells in the tumour environment, we performed Spearman’s correlation analysis on average immune cell densities across patients. Positive correlations were observed between CD3+CD8– and CD3+CD8+ T cells (R = 0.64, p<0.001), CD3+CD8– T cells and B cells (R = 0.71, p<0.001), and CD3+CD8+ T cells and B cells (R = 0.52, p<0.001) (Figure 5A). These findings suggest that crosstalk may occur between lymphocytes, which may prompt formation of niches within the PDAC microenvironment.

We next applied radius-based nearest neighbor clustering to ascertain potential interactions between immune cell types. With these, we identified three patterns of immune cell colocalization, which we termed “neighbourhoods”. The first was a macrophage dominant neighbourhood, composed primarily of CD16a+ M2-like macrophages, consistent with an immunosuppressive function (Figure 5B). The second was a T cell dominant neighbourhood, enriched for CD3+CD8– T cells, CD3+CD8+ T cells, and B cells, which may represent areas of organized adaptive immunity. The third was a disorganized neighbourhood, characterized by a heterogenous mix of infiltrating immune cells without a dominant cell type, suggestive of an unstructured (and perhaps inefficient) immune response.

The relative area represented by each neighbourhood varied between patients; however, in most tumours, the disorganized neighbourhood comprised the largest proportion of immune cells, followed by the macrophage dominant neighbourhood, and then the T cell dominant neighbourhood (Figures 5C, D). The disorganized and macrophage dominant neighbourhoods were observed in all patients, while the T cell dominant neighbourhood was present in ~90% of cases. Representative examples of each neighbourhood pattern are shown in Figure 5D.

To evaluate whether immune neighbourhoods were associated with patient outcomes, we performed Cox regression analyses comparing five-year OS in patients categorized as high or low for each neighbourhood. Having a higher density of the T cell dominant neighbourhood in the epithelium was associated with a decreased risk of death (HR: 0.41, 95% CI: 0.19-0.91, p<0.05). A higher proportion of the T cell dominant neighbourhood within the peritumoural region was associated with a decreased risk of death (HR: 0.45, 95% CI: 0.24-0.82, p<0.05) (Figure 5E). These associations were retained after multivariate Cox regression analysis (Supplementary Table S3). Neither the disorganized nor the macrophage dominant neighbourhoods were associated with OS. We conclude that the spatial co-localization of T cells and B cells proximal to the tumour is prognostically beneficial.